Model

Granite 4.1 8B Instruct

OPEN

IBM · Granite 4.1 family · released Apr 29, 2026

Successor to the Granite 4.0-H-Small MoE already in the corpus (dated 2025-10-02); IBM says the new 8B dense model matches or beats that prior 32B MoE flagship at 4x fewer parameters. Released alongside 3B and 30B siblings plus updated speech/vision/embedding/Guardian models.

ReasoningCodingVisionFunction callingTool useAgentic

965.1

Elo · rank #222

Parameters

Active params

8B (dense)

Context

512K tokens

Architecture

Dense decoder-only Transformer

License

Apache 2.0

Languages

—

API price (in/out)

No hosted API

Modalities

text · vision

Benchmark results

Bar shows position within the tracked field; marker = field best

AlpacaEvalHuman preference50.1%#19

best: Tülu 2+DPO 70B · 95.1%

Arena-HardHuman preference69.0%#35

best: Qwen3-235B-A22B (Non-Thinking) · 96.1%

Arena EloHuman preference1307#34

best: Claude Fable 5 · 1505

BIG-Bench HardReasoning80.5%#32

best: ERNIE 4.5 300B-A47B · 94.3%

BFCL v3Agents68.3%#14

best: Hunyuan-A13B · 78.3%

BigCodeBenchCoding35.0%#7

best: GPT-4o mini · 57.4%

GPQA DiamondReasoning42.0%#210

best: GPT-5.6 · 94.6%

GSM8KMath92.5%#26

best: Llama 3.1 405B · 96.8%

Humanity's Last ExamReasoning4.0%#129

best: Claude Sonnet 5 · 57.4%

HumanEval+Coding79.9%#15

best: Mistral Small 3.2 (24B) · 92.9%

HumanEvalCoding85.4%#42

best: Claude Opus 4.5 · 99.4%

IFBenchReasoning39.0%#48

best: MiniMax M3 · 83.0%

IFEvalReasoning87.1%#45

best: Gemma 4 26B A4B · 98.5%

MBPP+Coding73.8%#10

best: Llama 3.1 405B · 88.6%

MBPPCoding87.3%#2

best: Llama-3.3-Nemotron-Super-49B v1 (Reasoning On) · 91.3%

MGSMMath82.3%#26

best: OpenAI o4-mini · 93.7%

MMLU-ProKnowledge56.0%#120

best: Claude Fable 5 · 91.5%

MMLUKnowledge73.8%#128

best: OpenAI o3 · 92.9%

MMMLUKnowledge64.8%#48

best: Gemini 3.1 Pro · 92.6%

MT-BenchHuman preference8.61#21

best: Hunyuan-Large (A52B) · 9.4

SimpleQAKnowledge4.8%#45

best: GPT-4.5 · 62.5%

Terminal-Bench 2.0Coding3.0%#63

best: GPT-5.6 · 88.8%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Granite 4.1 family

Elo progression across releases

Granite 4.1 8B InstructApr 2026965.1

Resources

HFibm-granite/granite-4.1-8b DOCShttps://www.ibm.com/granite/docs/models/granite4-1 APIAvailable via IBM API

API price weights · each benchmark row carries its own source badge (see methodology)